Why Nitric stopped asking one AI agent to do everything

The team describes how frequent overnight pages, most of them noise, pushed them to build a workflow where specialized agents investigate incidents, track follow-up work, and prepare fixes while engineers review and correct the results. They introduce Sugabots, an open-source, self-hosted tool where people and agents share investigation history within team workspaces that connect to outside systems. Their Sentry triage agent chooses between paging the on-call engineer, scheduling an in-hours review, sending a small fix to an implementation agent, or archiving the issue, and a backlog agent records follow-up work in Linear. The central argument is that separating responsibilities lets a team judge each decision on its own, such as whether triage justified an interruption, whether a ticket kept enough evidence, and whether a proposed fix actually addressed the problem. The authors also stress checking for missed urgent incidents alongside fewer wake-up calls, since fewer interruptions could come from better classification or simply from fewer incidents. The piece closes by moving on to how each agent's tool permissions should match its particular role.